1604326b4SDaniel Borkmann // SPDX-License-Identifier: GPL-2.0 2604326b4SDaniel Borkmann /* Copyright (c) 2017 - 2018 Covalent IO, Inc. http://covalent.io */ 3604326b4SDaniel Borkmann 4604326b4SDaniel Borkmann #include <linux/skmsg.h> 5604326b4SDaniel Borkmann #include <linux/filter.h> 6604326b4SDaniel Borkmann #include <linux/bpf.h> 7604326b4SDaniel Borkmann #include <linux/init.h> 8604326b4SDaniel Borkmann #include <linux/wait.h> 9604326b4SDaniel Borkmann 10604326b4SDaniel Borkmann #include <net/inet_common.h> 110608c69cSJohn Fastabend #include <net/tls.h> 12604326b4SDaniel Borkmann 13604326b4SDaniel Borkmann static bool tcp_bpf_stream_read(const struct sock *sk) 14604326b4SDaniel Borkmann { 15604326b4SDaniel Borkmann struct sk_psock *psock; 16604326b4SDaniel Borkmann bool empty = true; 17604326b4SDaniel Borkmann 18604326b4SDaniel Borkmann rcu_read_lock(); 19604326b4SDaniel Borkmann psock = sk_psock(sk); 20604326b4SDaniel Borkmann if (likely(psock)) 21604326b4SDaniel Borkmann empty = list_empty(&psock->ingress_msg); 22604326b4SDaniel Borkmann rcu_read_unlock(); 23604326b4SDaniel Borkmann return !empty; 24604326b4SDaniel Borkmann } 25604326b4SDaniel Borkmann 26604326b4SDaniel Borkmann static int tcp_bpf_wait_data(struct sock *sk, struct sk_psock *psock, 27604326b4SDaniel Borkmann int flags, long timeo, int *err) 28604326b4SDaniel Borkmann { 29604326b4SDaniel Borkmann DEFINE_WAIT_FUNC(wait, woken_wake_function); 305fa2ca7cSJohn Fastabend int ret = 0; 315fa2ca7cSJohn Fastabend 325fa2ca7cSJohn Fastabend if (!timeo) 335fa2ca7cSJohn Fastabend return ret; 34604326b4SDaniel Borkmann 35604326b4SDaniel Borkmann add_wait_queue(sk_sleep(sk), &wait); 36604326b4SDaniel Borkmann sk_set_bit(SOCKWQ_ASYNC_WAITDATA, sk); 37604326b4SDaniel Borkmann ret = sk_wait_event(sk, &timeo, 38604326b4SDaniel Borkmann !list_empty(&psock->ingress_msg) || 39604326b4SDaniel Borkmann !skb_queue_empty(&sk->sk_receive_queue), &wait); 40604326b4SDaniel Borkmann sk_clear_bit(SOCKWQ_ASYNC_WAITDATA, sk); 41604326b4SDaniel Borkmann remove_wait_queue(sk_sleep(sk), &wait); 42604326b4SDaniel Borkmann return ret; 43604326b4SDaniel Borkmann } 44604326b4SDaniel Borkmann 45604326b4SDaniel Borkmann int __tcp_bpf_recvmsg(struct sock *sk, struct sk_psock *psock, 4602c558b2SJohn Fastabend struct msghdr *msg, int len, int flags) 47604326b4SDaniel Borkmann { 48604326b4SDaniel Borkmann struct iov_iter *iter = &msg->msg_iter; 4902c558b2SJohn Fastabend int peek = flags & MSG_PEEK; 50604326b4SDaniel Borkmann int i, ret, copied = 0; 51604326b4SDaniel Borkmann struct sk_msg *msg_rx; 52604326b4SDaniel Borkmann 53604326b4SDaniel Borkmann msg_rx = list_first_entry_or_null(&psock->ingress_msg, 54604326b4SDaniel Borkmann struct sk_msg, list); 5502c558b2SJohn Fastabend 5602c558b2SJohn Fastabend while (copied != len) { 5702c558b2SJohn Fastabend struct scatterlist *sge; 5802c558b2SJohn Fastabend 59604326b4SDaniel Borkmann if (unlikely(!msg_rx)) 60604326b4SDaniel Borkmann break; 61604326b4SDaniel Borkmann 62604326b4SDaniel Borkmann i = msg_rx->sg.start; 63604326b4SDaniel Borkmann do { 64604326b4SDaniel Borkmann struct page *page; 65604326b4SDaniel Borkmann int copy; 66604326b4SDaniel Borkmann 67604326b4SDaniel Borkmann sge = sk_msg_elem(msg_rx, i); 68604326b4SDaniel Borkmann copy = sge->length; 69604326b4SDaniel Borkmann page = sg_page(sge); 70604326b4SDaniel Borkmann if (copied + copy > len) 71604326b4SDaniel Borkmann copy = len - copied; 72604326b4SDaniel Borkmann ret = copy_page_to_iter(page, sge->offset, copy, iter); 73604326b4SDaniel Borkmann if (ret != copy) { 74604326b4SDaniel Borkmann msg_rx->sg.start = i; 75604326b4SDaniel Borkmann return -EFAULT; 76604326b4SDaniel Borkmann } 77604326b4SDaniel Borkmann 78604326b4SDaniel Borkmann copied += copy; 7902c558b2SJohn Fastabend if (likely(!peek)) { 80604326b4SDaniel Borkmann sge->offset += copy; 81604326b4SDaniel Borkmann sge->length -= copy; 82604326b4SDaniel Borkmann sk_mem_uncharge(sk, copy); 833f4c3127SJohn Fastabend msg_rx->sg.size -= copy; 8402c558b2SJohn Fastabend 85604326b4SDaniel Borkmann if (!sge->length) { 8602c558b2SJohn Fastabend sk_msg_iter_var_next(i); 87604326b4SDaniel Borkmann if (!msg_rx->skb) 88604326b4SDaniel Borkmann put_page(page); 89604326b4SDaniel Borkmann } 9002c558b2SJohn Fastabend } else { 9102c558b2SJohn Fastabend sk_msg_iter_var_next(i); 9202c558b2SJohn Fastabend } 93604326b4SDaniel Borkmann 94604326b4SDaniel Borkmann if (copied == len) 95604326b4SDaniel Borkmann break; 96604326b4SDaniel Borkmann } while (i != msg_rx->sg.end); 97604326b4SDaniel Borkmann 9802c558b2SJohn Fastabend if (unlikely(peek)) { 9902c558b2SJohn Fastabend msg_rx = list_next_entry(msg_rx, list); 10002c558b2SJohn Fastabend continue; 10102c558b2SJohn Fastabend } 10202c558b2SJohn Fastabend 103604326b4SDaniel Borkmann msg_rx->sg.start = i; 104604326b4SDaniel Borkmann if (!sge->length && msg_rx->sg.start == msg_rx->sg.end) { 105604326b4SDaniel Borkmann list_del(&msg_rx->list); 106604326b4SDaniel Borkmann if (msg_rx->skb) 107604326b4SDaniel Borkmann consume_skb(msg_rx->skb); 108604326b4SDaniel Borkmann kfree(msg_rx); 109604326b4SDaniel Borkmann } 11002c558b2SJohn Fastabend msg_rx = list_first_entry_or_null(&psock->ingress_msg, 11102c558b2SJohn Fastabend struct sk_msg, list); 112604326b4SDaniel Borkmann } 113604326b4SDaniel Borkmann 114604326b4SDaniel Borkmann return copied; 115604326b4SDaniel Borkmann } 116604326b4SDaniel Borkmann EXPORT_SYMBOL_GPL(__tcp_bpf_recvmsg); 117604326b4SDaniel Borkmann 118604326b4SDaniel Borkmann int tcp_bpf_recvmsg(struct sock *sk, struct msghdr *msg, size_t len, 119604326b4SDaniel Borkmann int nonblock, int flags, int *addr_len) 120604326b4SDaniel Borkmann { 121604326b4SDaniel Borkmann struct sk_psock *psock; 122604326b4SDaniel Borkmann int copied, ret; 123604326b4SDaniel Borkmann 124604326b4SDaniel Borkmann if (unlikely(flags & MSG_ERRQUEUE)) 125604326b4SDaniel Borkmann return inet_recv_error(sk, msg, len, addr_len); 126604326b4SDaniel Borkmann if (!skb_queue_empty(&sk->sk_receive_queue)) 127604326b4SDaniel Borkmann return tcp_recvmsg(sk, msg, len, nonblock, flags, addr_len); 128604326b4SDaniel Borkmann 129604326b4SDaniel Borkmann psock = sk_psock_get(sk); 130604326b4SDaniel Borkmann if (unlikely(!psock)) 131604326b4SDaniel Borkmann return tcp_recvmsg(sk, msg, len, nonblock, flags, addr_len); 132604326b4SDaniel Borkmann lock_sock(sk); 133604326b4SDaniel Borkmann msg_bytes_ready: 13402c558b2SJohn Fastabend copied = __tcp_bpf_recvmsg(sk, psock, msg, len, flags); 135604326b4SDaniel Borkmann if (!copied) { 136604326b4SDaniel Borkmann int data, err = 0; 137604326b4SDaniel Borkmann long timeo; 138604326b4SDaniel Borkmann 139604326b4SDaniel Borkmann timeo = sock_rcvtimeo(sk, nonblock); 140604326b4SDaniel Borkmann data = tcp_bpf_wait_data(sk, psock, flags, timeo, &err); 141604326b4SDaniel Borkmann if (data) { 142604326b4SDaniel Borkmann if (skb_queue_empty(&sk->sk_receive_queue)) 143604326b4SDaniel Borkmann goto msg_bytes_ready; 144604326b4SDaniel Borkmann release_sock(sk); 145604326b4SDaniel Borkmann sk_psock_put(sk, psock); 146604326b4SDaniel Borkmann return tcp_recvmsg(sk, msg, len, nonblock, flags, addr_len); 147604326b4SDaniel Borkmann } 148604326b4SDaniel Borkmann if (err) { 149604326b4SDaniel Borkmann ret = err; 150604326b4SDaniel Borkmann goto out; 151604326b4SDaniel Borkmann } 15227b31e68SJohn Fastabend copied = -EAGAIN; 153604326b4SDaniel Borkmann } 154604326b4SDaniel Borkmann ret = copied; 155604326b4SDaniel Borkmann out: 156604326b4SDaniel Borkmann release_sock(sk); 157604326b4SDaniel Borkmann sk_psock_put(sk, psock); 158604326b4SDaniel Borkmann return ret; 159604326b4SDaniel Borkmann } 160604326b4SDaniel Borkmann 161604326b4SDaniel Borkmann static int bpf_tcp_ingress(struct sock *sk, struct sk_psock *psock, 162604326b4SDaniel Borkmann struct sk_msg *msg, u32 apply_bytes, int flags) 163604326b4SDaniel Borkmann { 164604326b4SDaniel Borkmann bool apply = apply_bytes; 165604326b4SDaniel Borkmann struct scatterlist *sge; 166604326b4SDaniel Borkmann u32 size, copied = 0; 167604326b4SDaniel Borkmann struct sk_msg *tmp; 168604326b4SDaniel Borkmann int i, ret = 0; 169604326b4SDaniel Borkmann 170604326b4SDaniel Borkmann tmp = kzalloc(sizeof(*tmp), __GFP_NOWARN | GFP_KERNEL); 171604326b4SDaniel Borkmann if (unlikely(!tmp)) 172604326b4SDaniel Borkmann return -ENOMEM; 173604326b4SDaniel Borkmann 174604326b4SDaniel Borkmann lock_sock(sk); 175604326b4SDaniel Borkmann tmp->sg.start = msg->sg.start; 176604326b4SDaniel Borkmann i = msg->sg.start; 177604326b4SDaniel Borkmann do { 178604326b4SDaniel Borkmann sge = sk_msg_elem(msg, i); 179604326b4SDaniel Borkmann size = (apply && apply_bytes < sge->length) ? 180604326b4SDaniel Borkmann apply_bytes : sge->length; 181604326b4SDaniel Borkmann if (!sk_wmem_schedule(sk, size)) { 182604326b4SDaniel Borkmann if (!copied) 183604326b4SDaniel Borkmann ret = -ENOMEM; 184604326b4SDaniel Borkmann break; 185604326b4SDaniel Borkmann } 186604326b4SDaniel Borkmann 187604326b4SDaniel Borkmann sk_mem_charge(sk, size); 188604326b4SDaniel Borkmann sk_msg_xfer(tmp, msg, i, size); 189604326b4SDaniel Borkmann copied += size; 190604326b4SDaniel Borkmann if (sge->length) 191604326b4SDaniel Borkmann get_page(sk_msg_page(tmp, i)); 192604326b4SDaniel Borkmann sk_msg_iter_var_next(i); 193604326b4SDaniel Borkmann tmp->sg.end = i; 194604326b4SDaniel Borkmann if (apply) { 195604326b4SDaniel Borkmann apply_bytes -= size; 196604326b4SDaniel Borkmann if (!apply_bytes) 197604326b4SDaniel Borkmann break; 198604326b4SDaniel Borkmann } 199604326b4SDaniel Borkmann } while (i != msg->sg.end); 200604326b4SDaniel Borkmann 201604326b4SDaniel Borkmann if (!ret) { 202604326b4SDaniel Borkmann msg->sg.start = i; 203604326b4SDaniel Borkmann msg->sg.size -= apply_bytes; 204604326b4SDaniel Borkmann sk_psock_queue_msg(psock, tmp); 205552de910SJohn Fastabend sk_psock_data_ready(sk, psock); 206604326b4SDaniel Borkmann } else { 207604326b4SDaniel Borkmann sk_msg_free(sk, tmp); 208604326b4SDaniel Borkmann kfree(tmp); 209604326b4SDaniel Borkmann } 210604326b4SDaniel Borkmann 211604326b4SDaniel Borkmann release_sock(sk); 212604326b4SDaniel Borkmann return ret; 213604326b4SDaniel Borkmann } 214604326b4SDaniel Borkmann 215604326b4SDaniel Borkmann static int tcp_bpf_push(struct sock *sk, struct sk_msg *msg, u32 apply_bytes, 216604326b4SDaniel Borkmann int flags, bool uncharge) 217604326b4SDaniel Borkmann { 218604326b4SDaniel Borkmann bool apply = apply_bytes; 219604326b4SDaniel Borkmann struct scatterlist *sge; 220604326b4SDaniel Borkmann struct page *page; 221604326b4SDaniel Borkmann int size, ret = 0; 222604326b4SDaniel Borkmann u32 off; 223604326b4SDaniel Borkmann 224604326b4SDaniel Borkmann while (1) { 2250608c69cSJohn Fastabend bool has_tx_ulp; 2260608c69cSJohn Fastabend 227604326b4SDaniel Borkmann sge = sk_msg_elem(msg, msg->sg.start); 228604326b4SDaniel Borkmann size = (apply && apply_bytes < sge->length) ? 229604326b4SDaniel Borkmann apply_bytes : sge->length; 230604326b4SDaniel Borkmann off = sge->offset; 231604326b4SDaniel Borkmann page = sg_page(sge); 232604326b4SDaniel Borkmann 233604326b4SDaniel Borkmann tcp_rate_check_app_limited(sk); 234604326b4SDaniel Borkmann retry: 2350608c69cSJohn Fastabend has_tx_ulp = tls_sw_has_ctx_tx(sk); 2360608c69cSJohn Fastabend if (has_tx_ulp) { 2370608c69cSJohn Fastabend flags |= MSG_SENDPAGE_NOPOLICY; 2380608c69cSJohn Fastabend ret = kernel_sendpage_locked(sk, 2390608c69cSJohn Fastabend page, off, size, flags); 2400608c69cSJohn Fastabend } else { 241604326b4SDaniel Borkmann ret = do_tcp_sendpages(sk, page, off, size, flags); 2420608c69cSJohn Fastabend } 2430608c69cSJohn Fastabend 244604326b4SDaniel Borkmann if (ret <= 0) 245604326b4SDaniel Borkmann return ret; 246604326b4SDaniel Borkmann if (apply) 247604326b4SDaniel Borkmann apply_bytes -= ret; 248604326b4SDaniel Borkmann msg->sg.size -= ret; 249604326b4SDaniel Borkmann sge->offset += ret; 250604326b4SDaniel Borkmann sge->length -= ret; 251604326b4SDaniel Borkmann if (uncharge) 252604326b4SDaniel Borkmann sk_mem_uncharge(sk, ret); 253604326b4SDaniel Borkmann if (ret != size) { 254604326b4SDaniel Borkmann size -= ret; 255604326b4SDaniel Borkmann off += ret; 256604326b4SDaniel Borkmann goto retry; 257604326b4SDaniel Borkmann } 258604326b4SDaniel Borkmann if (!sge->length) { 259604326b4SDaniel Borkmann put_page(page); 260604326b4SDaniel Borkmann sk_msg_iter_next(msg, start); 261604326b4SDaniel Borkmann sg_init_table(sge, 1); 262604326b4SDaniel Borkmann if (msg->sg.start == msg->sg.end) 263604326b4SDaniel Borkmann break; 264604326b4SDaniel Borkmann } 265604326b4SDaniel Borkmann if (apply && !apply_bytes) 266604326b4SDaniel Borkmann break; 267604326b4SDaniel Borkmann } 268604326b4SDaniel Borkmann 269604326b4SDaniel Borkmann return 0; 270604326b4SDaniel Borkmann } 271604326b4SDaniel Borkmann 272604326b4SDaniel Borkmann static int tcp_bpf_push_locked(struct sock *sk, struct sk_msg *msg, 273604326b4SDaniel Borkmann u32 apply_bytes, int flags, bool uncharge) 274604326b4SDaniel Borkmann { 275604326b4SDaniel Borkmann int ret; 276604326b4SDaniel Borkmann 277604326b4SDaniel Borkmann lock_sock(sk); 278604326b4SDaniel Borkmann ret = tcp_bpf_push(sk, msg, apply_bytes, flags, uncharge); 279604326b4SDaniel Borkmann release_sock(sk); 280604326b4SDaniel Borkmann return ret; 281604326b4SDaniel Borkmann } 282604326b4SDaniel Borkmann 283604326b4SDaniel Borkmann int tcp_bpf_sendmsg_redir(struct sock *sk, struct sk_msg *msg, 284604326b4SDaniel Borkmann u32 bytes, int flags) 285604326b4SDaniel Borkmann { 286604326b4SDaniel Borkmann bool ingress = sk_msg_to_ingress(msg); 287604326b4SDaniel Borkmann struct sk_psock *psock = sk_psock_get(sk); 288604326b4SDaniel Borkmann int ret; 289604326b4SDaniel Borkmann 290604326b4SDaniel Borkmann if (unlikely(!psock)) { 291604326b4SDaniel Borkmann sk_msg_free(sk, msg); 292604326b4SDaniel Borkmann return 0; 293604326b4SDaniel Borkmann } 294604326b4SDaniel Borkmann ret = ingress ? bpf_tcp_ingress(sk, psock, msg, bytes, flags) : 295604326b4SDaniel Borkmann tcp_bpf_push_locked(sk, msg, bytes, flags, false); 296604326b4SDaniel Borkmann sk_psock_put(sk, psock); 297604326b4SDaniel Borkmann return ret; 298604326b4SDaniel Borkmann } 299604326b4SDaniel Borkmann EXPORT_SYMBOL_GPL(tcp_bpf_sendmsg_redir); 300604326b4SDaniel Borkmann 301604326b4SDaniel Borkmann static int tcp_bpf_send_verdict(struct sock *sk, struct sk_psock *psock, 302604326b4SDaniel Borkmann struct sk_msg *msg, int *copied, int flags) 303604326b4SDaniel Borkmann { 304604326b4SDaniel Borkmann bool cork = false, enospc = msg->sg.start == msg->sg.end; 305604326b4SDaniel Borkmann struct sock *sk_redir; 3067246d8edSJohn Fastabend u32 tosend, delta = 0; 307604326b4SDaniel Borkmann int ret; 308604326b4SDaniel Borkmann 309604326b4SDaniel Borkmann more_data: 3107246d8edSJohn Fastabend if (psock->eval == __SK_NONE) { 3117246d8edSJohn Fastabend /* Track delta in msg size to add/subtract it on SK_DROP from 3127246d8edSJohn Fastabend * returned to user copied size. This ensures user doesn't 3137246d8edSJohn Fastabend * get a positive return code with msg_cut_data and SK_DROP 3147246d8edSJohn Fastabend * verdict. 3157246d8edSJohn Fastabend */ 3167246d8edSJohn Fastabend delta = msg->sg.size; 317604326b4SDaniel Borkmann psock->eval = sk_psock_msg_verdict(sk, psock, msg); 3187246d8edSJohn Fastabend if (msg->sg.size < delta) 3197246d8edSJohn Fastabend delta -= msg->sg.size; 3207246d8edSJohn Fastabend else 3217246d8edSJohn Fastabend delta = 0; 3227246d8edSJohn Fastabend } 323604326b4SDaniel Borkmann 324604326b4SDaniel Borkmann if (msg->cork_bytes && 325604326b4SDaniel Borkmann msg->cork_bytes > msg->sg.size && !enospc) { 326604326b4SDaniel Borkmann psock->cork_bytes = msg->cork_bytes - msg->sg.size; 327604326b4SDaniel Borkmann if (!psock->cork) { 328604326b4SDaniel Borkmann psock->cork = kzalloc(sizeof(*psock->cork), 329604326b4SDaniel Borkmann GFP_ATOMIC | __GFP_NOWARN); 330604326b4SDaniel Borkmann if (!psock->cork) 331604326b4SDaniel Borkmann return -ENOMEM; 332604326b4SDaniel Borkmann } 333604326b4SDaniel Borkmann memcpy(psock->cork, msg, sizeof(*msg)); 334604326b4SDaniel Borkmann return 0; 335604326b4SDaniel Borkmann } 336604326b4SDaniel Borkmann 337604326b4SDaniel Borkmann tosend = msg->sg.size; 338604326b4SDaniel Borkmann if (psock->apply_bytes && psock->apply_bytes < tosend) 339604326b4SDaniel Borkmann tosend = psock->apply_bytes; 340604326b4SDaniel Borkmann 341604326b4SDaniel Borkmann switch (psock->eval) { 342604326b4SDaniel Borkmann case __SK_PASS: 343604326b4SDaniel Borkmann ret = tcp_bpf_push(sk, msg, tosend, flags, true); 344604326b4SDaniel Borkmann if (unlikely(ret)) { 345604326b4SDaniel Borkmann *copied -= sk_msg_free(sk, msg); 346604326b4SDaniel Borkmann break; 347604326b4SDaniel Borkmann } 348604326b4SDaniel Borkmann sk_msg_apply_bytes(psock, tosend); 349604326b4SDaniel Borkmann break; 350604326b4SDaniel Borkmann case __SK_REDIRECT: 351604326b4SDaniel Borkmann sk_redir = psock->sk_redir; 352604326b4SDaniel Borkmann sk_msg_apply_bytes(psock, tosend); 353604326b4SDaniel Borkmann if (psock->cork) { 354604326b4SDaniel Borkmann cork = true; 355604326b4SDaniel Borkmann psock->cork = NULL; 356604326b4SDaniel Borkmann } 357604326b4SDaniel Borkmann sk_msg_return(sk, msg, tosend); 358604326b4SDaniel Borkmann release_sock(sk); 359604326b4SDaniel Borkmann ret = tcp_bpf_sendmsg_redir(sk_redir, msg, tosend, flags); 360604326b4SDaniel Borkmann lock_sock(sk); 361604326b4SDaniel Borkmann if (unlikely(ret < 0)) { 362604326b4SDaniel Borkmann int free = sk_msg_free_nocharge(sk, msg); 363604326b4SDaniel Borkmann 364604326b4SDaniel Borkmann if (!cork) 365604326b4SDaniel Borkmann *copied -= free; 366604326b4SDaniel Borkmann } 367604326b4SDaniel Borkmann if (cork) { 368604326b4SDaniel Borkmann sk_msg_free(sk, msg); 369604326b4SDaniel Borkmann kfree(msg); 370604326b4SDaniel Borkmann msg = NULL; 371604326b4SDaniel Borkmann ret = 0; 372604326b4SDaniel Borkmann } 373604326b4SDaniel Borkmann break; 374604326b4SDaniel Borkmann case __SK_DROP: 375604326b4SDaniel Borkmann default: 376604326b4SDaniel Borkmann sk_msg_free_partial(sk, msg, tosend); 377604326b4SDaniel Borkmann sk_msg_apply_bytes(psock, tosend); 3787246d8edSJohn Fastabend *copied -= (tosend + delta); 379604326b4SDaniel Borkmann return -EACCES; 380604326b4SDaniel Borkmann } 381604326b4SDaniel Borkmann 382604326b4SDaniel Borkmann if (likely(!ret)) { 383604326b4SDaniel Borkmann if (!psock->apply_bytes) { 384604326b4SDaniel Borkmann psock->eval = __SK_NONE; 385604326b4SDaniel Borkmann if (psock->sk_redir) { 386604326b4SDaniel Borkmann sock_put(psock->sk_redir); 387604326b4SDaniel Borkmann psock->sk_redir = NULL; 388604326b4SDaniel Borkmann } 389604326b4SDaniel Borkmann } 390604326b4SDaniel Borkmann if (msg && 391604326b4SDaniel Borkmann msg->sg.data[msg->sg.start].page_link && 392604326b4SDaniel Borkmann msg->sg.data[msg->sg.start].length) 393604326b4SDaniel Borkmann goto more_data; 394604326b4SDaniel Borkmann } 395604326b4SDaniel Borkmann return ret; 396604326b4SDaniel Borkmann } 397604326b4SDaniel Borkmann 398604326b4SDaniel Borkmann static int tcp_bpf_sendmsg(struct sock *sk, struct msghdr *msg, size_t size) 399604326b4SDaniel Borkmann { 400604326b4SDaniel Borkmann struct sk_msg tmp, *msg_tx = NULL; 401604326b4SDaniel Borkmann int copied = 0, err = 0; 402604326b4SDaniel Borkmann struct sk_psock *psock; 403604326b4SDaniel Borkmann long timeo; 404*41477662SJakub Kicinski int flags; 405*41477662SJakub Kicinski 406*41477662SJakub Kicinski /* Don't let internal do_tcp_sendpages() flags through */ 407*41477662SJakub Kicinski flags = (msg->msg_flags & ~MSG_SENDPAGE_DECRYPTED); 408*41477662SJakub Kicinski flags |= MSG_NO_SHARED_FRAGS; 409604326b4SDaniel Borkmann 410604326b4SDaniel Borkmann psock = sk_psock_get(sk); 411604326b4SDaniel Borkmann if (unlikely(!psock)) 412604326b4SDaniel Borkmann return tcp_sendmsg(sk, msg, size); 413604326b4SDaniel Borkmann 414604326b4SDaniel Borkmann lock_sock(sk); 415604326b4SDaniel Borkmann timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT); 416604326b4SDaniel Borkmann while (msg_data_left(msg)) { 417604326b4SDaniel Borkmann bool enospc = false; 418604326b4SDaniel Borkmann u32 copy, osize; 419604326b4SDaniel Borkmann 420604326b4SDaniel Borkmann if (sk->sk_err) { 421604326b4SDaniel Borkmann err = -sk->sk_err; 422604326b4SDaniel Borkmann goto out_err; 423604326b4SDaniel Borkmann } 424604326b4SDaniel Borkmann 425604326b4SDaniel Borkmann copy = msg_data_left(msg); 426604326b4SDaniel Borkmann if (!sk_stream_memory_free(sk)) 427604326b4SDaniel Borkmann goto wait_for_sndbuf; 428604326b4SDaniel Borkmann if (psock->cork) { 429604326b4SDaniel Borkmann msg_tx = psock->cork; 430604326b4SDaniel Borkmann } else { 431604326b4SDaniel Borkmann msg_tx = &tmp; 432604326b4SDaniel Borkmann sk_msg_init(msg_tx); 433604326b4SDaniel Borkmann } 434604326b4SDaniel Borkmann 435604326b4SDaniel Borkmann osize = msg_tx->sg.size; 436604326b4SDaniel Borkmann err = sk_msg_alloc(sk, msg_tx, msg_tx->sg.size + copy, msg_tx->sg.end - 1); 437604326b4SDaniel Borkmann if (err) { 438604326b4SDaniel Borkmann if (err != -ENOSPC) 439604326b4SDaniel Borkmann goto wait_for_memory; 440604326b4SDaniel Borkmann enospc = true; 441604326b4SDaniel Borkmann copy = msg_tx->sg.size - osize; 442604326b4SDaniel Borkmann } 443604326b4SDaniel Borkmann 444604326b4SDaniel Borkmann err = sk_msg_memcopy_from_iter(sk, &msg->msg_iter, msg_tx, 445604326b4SDaniel Borkmann copy); 446604326b4SDaniel Borkmann if (err < 0) { 447604326b4SDaniel Borkmann sk_msg_trim(sk, msg_tx, osize); 448604326b4SDaniel Borkmann goto out_err; 449604326b4SDaniel Borkmann } 450604326b4SDaniel Borkmann 451604326b4SDaniel Borkmann copied += copy; 452604326b4SDaniel Borkmann if (psock->cork_bytes) { 453604326b4SDaniel Borkmann if (size > psock->cork_bytes) 454604326b4SDaniel Borkmann psock->cork_bytes = 0; 455604326b4SDaniel Borkmann else 456604326b4SDaniel Borkmann psock->cork_bytes -= size; 457604326b4SDaniel Borkmann if (psock->cork_bytes && !enospc) 458604326b4SDaniel Borkmann goto out_err; 459604326b4SDaniel Borkmann /* All cork bytes are accounted, rerun the prog. */ 460604326b4SDaniel Borkmann psock->eval = __SK_NONE; 461604326b4SDaniel Borkmann psock->cork_bytes = 0; 462604326b4SDaniel Borkmann } 463604326b4SDaniel Borkmann 464604326b4SDaniel Borkmann err = tcp_bpf_send_verdict(sk, psock, msg_tx, &copied, flags); 465604326b4SDaniel Borkmann if (unlikely(err < 0)) 466604326b4SDaniel Borkmann goto out_err; 467604326b4SDaniel Borkmann continue; 468604326b4SDaniel Borkmann wait_for_sndbuf: 469604326b4SDaniel Borkmann set_bit(SOCK_NOSPACE, &sk->sk_socket->flags); 470604326b4SDaniel Borkmann wait_for_memory: 471604326b4SDaniel Borkmann err = sk_stream_wait_memory(sk, &timeo); 472604326b4SDaniel Borkmann if (err) { 473604326b4SDaniel Borkmann if (msg_tx && msg_tx != psock->cork) 474604326b4SDaniel Borkmann sk_msg_free(sk, msg_tx); 475604326b4SDaniel Borkmann goto out_err; 476604326b4SDaniel Borkmann } 477604326b4SDaniel Borkmann } 478604326b4SDaniel Borkmann out_err: 479604326b4SDaniel Borkmann if (err < 0) 480604326b4SDaniel Borkmann err = sk_stream_error(sk, msg->msg_flags, err); 481604326b4SDaniel Borkmann release_sock(sk); 482604326b4SDaniel Borkmann sk_psock_put(sk, psock); 483604326b4SDaniel Borkmann return copied ? copied : err; 484604326b4SDaniel Borkmann } 485604326b4SDaniel Borkmann 486604326b4SDaniel Borkmann static int tcp_bpf_sendpage(struct sock *sk, struct page *page, int offset, 487604326b4SDaniel Borkmann size_t size, int flags) 488604326b4SDaniel Borkmann { 489604326b4SDaniel Borkmann struct sk_msg tmp, *msg = NULL; 490604326b4SDaniel Borkmann int err = 0, copied = 0; 491604326b4SDaniel Borkmann struct sk_psock *psock; 492604326b4SDaniel Borkmann bool enospc = false; 493604326b4SDaniel Borkmann 494604326b4SDaniel Borkmann psock = sk_psock_get(sk); 495604326b4SDaniel Borkmann if (unlikely(!psock)) 496604326b4SDaniel Borkmann return tcp_sendpage(sk, page, offset, size, flags); 497604326b4SDaniel Borkmann 498604326b4SDaniel Borkmann lock_sock(sk); 499604326b4SDaniel Borkmann if (psock->cork) { 500604326b4SDaniel Borkmann msg = psock->cork; 501604326b4SDaniel Borkmann } else { 502604326b4SDaniel Borkmann msg = &tmp; 503604326b4SDaniel Borkmann sk_msg_init(msg); 504604326b4SDaniel Borkmann } 505604326b4SDaniel Borkmann 506604326b4SDaniel Borkmann /* Catch case where ring is full and sendpage is stalled. */ 507604326b4SDaniel Borkmann if (unlikely(sk_msg_full(msg))) 508604326b4SDaniel Borkmann goto out_err; 509604326b4SDaniel Borkmann 510604326b4SDaniel Borkmann sk_msg_page_add(msg, page, size, offset); 511604326b4SDaniel Borkmann sk_mem_charge(sk, size); 512604326b4SDaniel Borkmann copied = size; 513604326b4SDaniel Borkmann if (sk_msg_full(msg)) 514604326b4SDaniel Borkmann enospc = true; 515604326b4SDaniel Borkmann if (psock->cork_bytes) { 516604326b4SDaniel Borkmann if (size > psock->cork_bytes) 517604326b4SDaniel Borkmann psock->cork_bytes = 0; 518604326b4SDaniel Borkmann else 519604326b4SDaniel Borkmann psock->cork_bytes -= size; 520604326b4SDaniel Borkmann if (psock->cork_bytes && !enospc) 521604326b4SDaniel Borkmann goto out_err; 522604326b4SDaniel Borkmann /* All cork bytes are accounted, rerun the prog. */ 523604326b4SDaniel Borkmann psock->eval = __SK_NONE; 524604326b4SDaniel Borkmann psock->cork_bytes = 0; 525604326b4SDaniel Borkmann } 526604326b4SDaniel Borkmann 527604326b4SDaniel Borkmann err = tcp_bpf_send_verdict(sk, psock, msg, &copied, flags); 528604326b4SDaniel Borkmann out_err: 529604326b4SDaniel Borkmann release_sock(sk); 530604326b4SDaniel Borkmann sk_psock_put(sk, psock); 531604326b4SDaniel Borkmann return copied ? copied : err; 532604326b4SDaniel Borkmann } 533604326b4SDaniel Borkmann 534604326b4SDaniel Borkmann static void tcp_bpf_remove(struct sock *sk, struct sk_psock *psock) 535604326b4SDaniel Borkmann { 536604326b4SDaniel Borkmann struct sk_psock_link *link; 537604326b4SDaniel Borkmann 538604326b4SDaniel Borkmann while ((link = sk_psock_link_pop(psock))) { 539604326b4SDaniel Borkmann sk_psock_unlink(sk, link); 540604326b4SDaniel Borkmann sk_psock_free_link(link); 541604326b4SDaniel Borkmann } 542604326b4SDaniel Borkmann } 543604326b4SDaniel Borkmann 544604326b4SDaniel Borkmann static void tcp_bpf_unhash(struct sock *sk) 545604326b4SDaniel Borkmann { 546604326b4SDaniel Borkmann void (*saved_unhash)(struct sock *sk); 547604326b4SDaniel Borkmann struct sk_psock *psock; 548604326b4SDaniel Borkmann 549604326b4SDaniel Borkmann rcu_read_lock(); 550604326b4SDaniel Borkmann psock = sk_psock(sk); 551604326b4SDaniel Borkmann if (unlikely(!psock)) { 552604326b4SDaniel Borkmann rcu_read_unlock(); 553604326b4SDaniel Borkmann if (sk->sk_prot->unhash) 554604326b4SDaniel Borkmann sk->sk_prot->unhash(sk); 555604326b4SDaniel Borkmann return; 556604326b4SDaniel Borkmann } 557604326b4SDaniel Borkmann 558604326b4SDaniel Borkmann saved_unhash = psock->saved_unhash; 559604326b4SDaniel Borkmann tcp_bpf_remove(sk, psock); 560604326b4SDaniel Borkmann rcu_read_unlock(); 561604326b4SDaniel Borkmann saved_unhash(sk); 562604326b4SDaniel Borkmann } 563604326b4SDaniel Borkmann 564604326b4SDaniel Borkmann static void tcp_bpf_close(struct sock *sk, long timeout) 565604326b4SDaniel Borkmann { 566604326b4SDaniel Borkmann void (*saved_close)(struct sock *sk, long timeout); 567604326b4SDaniel Borkmann struct sk_psock *psock; 568604326b4SDaniel Borkmann 569604326b4SDaniel Borkmann lock_sock(sk); 570604326b4SDaniel Borkmann rcu_read_lock(); 571604326b4SDaniel Borkmann psock = sk_psock(sk); 572604326b4SDaniel Borkmann if (unlikely(!psock)) { 573604326b4SDaniel Borkmann rcu_read_unlock(); 574604326b4SDaniel Borkmann release_sock(sk); 575604326b4SDaniel Borkmann return sk->sk_prot->close(sk, timeout); 576604326b4SDaniel Borkmann } 577604326b4SDaniel Borkmann 578604326b4SDaniel Borkmann saved_close = psock->saved_close; 579604326b4SDaniel Borkmann tcp_bpf_remove(sk, psock); 580604326b4SDaniel Borkmann rcu_read_unlock(); 581604326b4SDaniel Borkmann release_sock(sk); 582604326b4SDaniel Borkmann saved_close(sk, timeout); 583604326b4SDaniel Borkmann } 584604326b4SDaniel Borkmann 585604326b4SDaniel Borkmann enum { 586604326b4SDaniel Borkmann TCP_BPF_IPV4, 587604326b4SDaniel Borkmann TCP_BPF_IPV6, 588604326b4SDaniel Borkmann TCP_BPF_NUM_PROTS, 589604326b4SDaniel Borkmann }; 590604326b4SDaniel Borkmann 591604326b4SDaniel Borkmann enum { 592604326b4SDaniel Borkmann TCP_BPF_BASE, 593604326b4SDaniel Borkmann TCP_BPF_TX, 594604326b4SDaniel Borkmann TCP_BPF_NUM_CFGS, 595604326b4SDaniel Borkmann }; 596604326b4SDaniel Borkmann 597604326b4SDaniel Borkmann static struct proto *tcpv6_prot_saved __read_mostly; 598604326b4SDaniel Borkmann static DEFINE_SPINLOCK(tcpv6_prot_lock); 599604326b4SDaniel Borkmann static struct proto tcp_bpf_prots[TCP_BPF_NUM_PROTS][TCP_BPF_NUM_CFGS]; 600604326b4SDaniel Borkmann 601604326b4SDaniel Borkmann static void tcp_bpf_rebuild_protos(struct proto prot[TCP_BPF_NUM_CFGS], 602604326b4SDaniel Borkmann struct proto *base) 603604326b4SDaniel Borkmann { 604604326b4SDaniel Borkmann prot[TCP_BPF_BASE] = *base; 605604326b4SDaniel Borkmann prot[TCP_BPF_BASE].unhash = tcp_bpf_unhash; 606604326b4SDaniel Borkmann prot[TCP_BPF_BASE].close = tcp_bpf_close; 607604326b4SDaniel Borkmann prot[TCP_BPF_BASE].recvmsg = tcp_bpf_recvmsg; 608604326b4SDaniel Borkmann prot[TCP_BPF_BASE].stream_memory_read = tcp_bpf_stream_read; 609604326b4SDaniel Borkmann 610604326b4SDaniel Borkmann prot[TCP_BPF_TX] = prot[TCP_BPF_BASE]; 611604326b4SDaniel Borkmann prot[TCP_BPF_TX].sendmsg = tcp_bpf_sendmsg; 612604326b4SDaniel Borkmann prot[TCP_BPF_TX].sendpage = tcp_bpf_sendpage; 613604326b4SDaniel Borkmann } 614604326b4SDaniel Borkmann 615604326b4SDaniel Borkmann static void tcp_bpf_check_v6_needs_rebuild(struct sock *sk, struct proto *ops) 616604326b4SDaniel Borkmann { 617604326b4SDaniel Borkmann if (sk->sk_family == AF_INET6 && 618604326b4SDaniel Borkmann unlikely(ops != smp_load_acquire(&tcpv6_prot_saved))) { 619604326b4SDaniel Borkmann spin_lock_bh(&tcpv6_prot_lock); 620604326b4SDaniel Borkmann if (likely(ops != tcpv6_prot_saved)) { 621604326b4SDaniel Borkmann tcp_bpf_rebuild_protos(tcp_bpf_prots[TCP_BPF_IPV6], ops); 622604326b4SDaniel Borkmann smp_store_release(&tcpv6_prot_saved, ops); 623604326b4SDaniel Borkmann } 624604326b4SDaniel Borkmann spin_unlock_bh(&tcpv6_prot_lock); 625604326b4SDaniel Borkmann } 626604326b4SDaniel Borkmann } 627604326b4SDaniel Borkmann 628604326b4SDaniel Borkmann static int __init tcp_bpf_v4_build_proto(void) 629604326b4SDaniel Borkmann { 630604326b4SDaniel Borkmann tcp_bpf_rebuild_protos(tcp_bpf_prots[TCP_BPF_IPV4], &tcp_prot); 631604326b4SDaniel Borkmann return 0; 632604326b4SDaniel Borkmann } 633604326b4SDaniel Borkmann core_initcall(tcp_bpf_v4_build_proto); 634604326b4SDaniel Borkmann 635604326b4SDaniel Borkmann static void tcp_bpf_update_sk_prot(struct sock *sk, struct sk_psock *psock) 636604326b4SDaniel Borkmann { 637604326b4SDaniel Borkmann int family = sk->sk_family == AF_INET6 ? TCP_BPF_IPV6 : TCP_BPF_IPV4; 638604326b4SDaniel Borkmann int config = psock->progs.msg_parser ? TCP_BPF_TX : TCP_BPF_BASE; 639604326b4SDaniel Borkmann 640604326b4SDaniel Borkmann sk_psock_update_proto(sk, psock, &tcp_bpf_prots[family][config]); 641604326b4SDaniel Borkmann } 642604326b4SDaniel Borkmann 643604326b4SDaniel Borkmann static void tcp_bpf_reinit_sk_prot(struct sock *sk, struct sk_psock *psock) 644604326b4SDaniel Borkmann { 645604326b4SDaniel Borkmann int family = sk->sk_family == AF_INET6 ? TCP_BPF_IPV6 : TCP_BPF_IPV4; 646604326b4SDaniel Borkmann int config = psock->progs.msg_parser ? TCP_BPF_TX : TCP_BPF_BASE; 647604326b4SDaniel Borkmann 648604326b4SDaniel Borkmann /* Reinit occurs when program types change e.g. TCP_BPF_TX is removed 649604326b4SDaniel Borkmann * or added requiring sk_prot hook updates. We keep original saved 650604326b4SDaniel Borkmann * hooks in this case. 651604326b4SDaniel Borkmann */ 652604326b4SDaniel Borkmann sk->sk_prot = &tcp_bpf_prots[family][config]; 653604326b4SDaniel Borkmann } 654604326b4SDaniel Borkmann 655604326b4SDaniel Borkmann static int tcp_bpf_assert_proto_ops(struct proto *ops) 656604326b4SDaniel Borkmann { 657604326b4SDaniel Borkmann /* In order to avoid retpoline, we make assumptions when we call 658604326b4SDaniel Borkmann * into ops if e.g. a psock is not present. Make sure they are 659604326b4SDaniel Borkmann * indeed valid assumptions. 660604326b4SDaniel Borkmann */ 661604326b4SDaniel Borkmann return ops->recvmsg == tcp_recvmsg && 662604326b4SDaniel Borkmann ops->sendmsg == tcp_sendmsg && 663604326b4SDaniel Borkmann ops->sendpage == tcp_sendpage ? 0 : -ENOTSUPP; 664604326b4SDaniel Borkmann } 665604326b4SDaniel Borkmann 666604326b4SDaniel Borkmann void tcp_bpf_reinit(struct sock *sk) 667604326b4SDaniel Borkmann { 668604326b4SDaniel Borkmann struct sk_psock *psock; 669604326b4SDaniel Borkmann 670604326b4SDaniel Borkmann sock_owned_by_me(sk); 671604326b4SDaniel Borkmann 672604326b4SDaniel Borkmann rcu_read_lock(); 673604326b4SDaniel Borkmann psock = sk_psock(sk); 674604326b4SDaniel Borkmann tcp_bpf_reinit_sk_prot(sk, psock); 675604326b4SDaniel Borkmann rcu_read_unlock(); 676604326b4SDaniel Borkmann } 677604326b4SDaniel Borkmann 678604326b4SDaniel Borkmann int tcp_bpf_init(struct sock *sk) 679604326b4SDaniel Borkmann { 680604326b4SDaniel Borkmann struct proto *ops = READ_ONCE(sk->sk_prot); 681604326b4SDaniel Borkmann struct sk_psock *psock; 682604326b4SDaniel Borkmann 683604326b4SDaniel Borkmann sock_owned_by_me(sk); 684604326b4SDaniel Borkmann 685604326b4SDaniel Borkmann rcu_read_lock(); 686604326b4SDaniel Borkmann psock = sk_psock(sk); 687604326b4SDaniel Borkmann if (unlikely(!psock || psock->sk_proto || 688604326b4SDaniel Borkmann tcp_bpf_assert_proto_ops(ops))) { 689604326b4SDaniel Borkmann rcu_read_unlock(); 690604326b4SDaniel Borkmann return -EINVAL; 691604326b4SDaniel Borkmann } 692604326b4SDaniel Borkmann tcp_bpf_check_v6_needs_rebuild(sk, ops); 693604326b4SDaniel Borkmann tcp_bpf_update_sk_prot(sk, psock); 694604326b4SDaniel Borkmann rcu_read_unlock(); 695604326b4SDaniel Borkmann return 0; 696604326b4SDaniel Borkmann } 697